مدل‌سازی دوبعدی بی‌هنجاری‌های مغناطیسی با استفاده از شبکه عصبی پیشخور

نوع مقاله: مقاله تحقیقی‌ (پژوهشی‌)

نویسندگان

1 دانشکده مهندسی معدن، پردیس دانشکده‌های فنی دانشگاه تهران، ایران

2 مؤسسه ژئوفیزیک دانشگاه تهران، ایران

3 دانشکده مهندسی برق و کامپیوتر، پردیس دانشکده‌های فنی دانشگاه تهران، ایران

چکیده

دراین مقاله، برای مدل‌سازی بی‌هنجاری‌های مغناطیسی از شبکه عصبی پیشخور استفاده شده، و مدل‌سازی با فرض شکل دایک شیبدار با گسترش نامحدود، انجام شده است. این روش قابلیت تخمین تمام پارامترهای هندسی یعنی؛ مختصات مرکز دایک بر روی پروفیل، عمق، شیب و عرض دایک را دارد.
ابتدا کارائی این روش، با مدل‌های مصنوعی بدون نوفه و نوفه‌دار آزمایش شد، که نتایج رضایت بخشی بدست آمد. سپس از آن برای تفسیر داده‌های مغناطیس‌‌سنجی معدن مروارید زنجان، استفاده شد. نتایج حاصل از مدل‌سازی، با نتایج روش دیکانولوشن اولر و اطلاعات حاصل از ترانشه‌ها و حفاری مطابقت بسیار خوبی دارد.

کلیدواژه‌ها


عنوان مقاله [English]

2D modeling of magnetic anomaly through feed forward neural network

نویسندگان [English]

  • Ahmad Afshar 1
  • Meysam Abedi 1
  • Gholam Hossain Norouzi 1
  • Vahid Ebrahimzadeh Ardestani 2
  • Caro Lucas 3
چکیده [English]

One of the most important goals of magnetic data interpretation is to determine location, depth and shape of the magnetic anomaly. Extensive use of magnetic surveys in the field of mineral exploration, geology and environmental application make earth scientists to present suitable interpretation schemes.
Neural networks are part of a much wider field called artificial intelligence which have computer algorithm that solve several types of problems. The problems include classification, parameter estimation, parameter prediction, pattern recognition, completion, association, filtering, and optimization. The NNs are used in different aspects of interpretation and modeling of geophysics data. They are used for inverting geophysical data involving problems for which no easy solutions exist.
In this paper, Feed Forward Neural Network (FNN) is used for the magnetic anomaly modeling with assumption an infinite depth extent dike. The method can estimate all the geometric parameters of the dike; horizontal location on profile, width, depth and dip.
We used a three layer FNN; consisting of 11 neurons in the input layer, 20 neurons in the hidden layer and 4 neurons in the output layer. For the first and hidden layers, and the last layer sigmoid and linear activation functions are used, respectively. Here, the horizontal location, width, depth and dip of the dike were defined as the output and the magnetic profile data as the input of the neural network. The training of the network was done through synthetic data which were produced by forward modeling.
For preventing the neural network from holding to any particular sequence, the input data were assigned in a random and then, the profile related to these data was obtained by dike equation. For the FNN to recognize the pattern of the profile data, some parameters are defined as the input of the FNN. These parameters should be separated. In other words, they must have a relation with the geometrical parameters of the dike. In definition of the parameters from the Anomaly curve, maximum and minimum points, width of curve in points of 75% and 50% of maximum, area of positive and negative parts of anomaly curve are used. Furthermore, the width and depth of the dike may be found through using a horizontal derivative of anomaly and a derivative of Hilbert transform. Horizontal derivation of Anomaly curve and Hilbert transform of horizontal derivation, especially from their intersection points, are also used as input parameters of neural network.
The validity of this method was tested by using of noise-free and noise-corrupted synthetic models that satisfactory results were obtained. Then the Morvarid mine ,which located at a distance of 30Km south east of Zanjan, near Aliabad village, is chosen as a real data application. The outputs show a good accordance with the Euler method and the tranches results.
The FNN inversion provides satisfactory results despite the noise. Various techniques are applied for interpretation of magnetic data. Most of them estimate only the depth. So, careful determination of dip, depth and width of the dike are the benefits of the use of the FNN.

کلیدواژه‌ها [English]

  • Feed forward neural network
  • magnetic anomaly
  • Dike
  • training data